Enhancing Brain Tumor Classification Accuracy through Multi-Modal MRI Analysis and Advanced Ensemble Deep Learning

This paper focused on the multi-class brain tumor classification problem, using deep learning methodologies based on magnetic resonance imaging (MRI) images. Two state-of-the-art deep learning models, EfficientNet-B3 and ConvNeXt-Tiny, were fine-tuned through transfer learning with custom classifier heads, and used a publicly available dataset from Kaggle, leveraging various data augmentation techniques to help reduce class imbalance. These two models were trained to the same hyperparameters and had the same metrics applied to the evaluation. Both models were assessed using Accuracy, F1-Score, Recall, and Precision. EfficientNet-B3 achieved an accuracy of 96.85% and ConvNeXt-Tiny achieved 95.88%. After observing the strengths of both models, a weighted ensemble (60% EfficientNet-B3 and 40% ConvNeXt-Tiny) was created, achieving a final accuracy of 97.07%. Based on the results of the study, it demonstrates how deep learning and ensemble techniques can improve diagnostic accuracy in a complex multi-class brain tumor classification task.